MicrostockGroup

Racephoto's comparison settled the issue for me.  I was surprised by the difference in the two images. 

Gonna keep saving at 12.

If you dig into the JPG compression algorithm, you will see that the information reduction is done by combining slightly different adjacent pixels into monochrome squares > 1px. When there is a lot of noisy info in the image, like in trees or grass, the size reduction is much smallerr than with clouds or scenes with many slow gradients, like cars.

One of the JPG quality parameters is the size of the squares.

It follows that you can see a clear effect in slow gradients like skies and clouds (where you can easily see the JPG square artifacts in medium-blue clouds) and on high-contrast luminance edges like dark suits on an overwhite. JPG is bad handling graphics and hence the edges on overwhites, which will show up with more jitter around the edge and into the white area near the border.
This can cause problems for a designer-buyer that wants to re-extract the object from the white.

Conclusion: q12 for overwhites and rasterized graphics, q10 for the rest.